Detectron2.1

This repository contains experiments on the generation of Adversarial Examples for Object Detection tasks, in particular for logo and input box detection for phishing website detection applications.

The primary technique used is the Dense Adversary Generation (DAG) algorithm from Adversarial Examples for Semantic Segmentation and Object Detection (Xie et al., ICCV 2017), and is implemented on the Detectron2 framework.

Report on Adversarial Examples for Object Detection: REPORT.md

Installation

git clone https://github.com/lemonwaffle/detectron2-1.git

To install Detectron2 and its dependencies, refer to the official installation instructions.

Directory

Make sure the data files are organized as follows:

├── GETTING_STARTED.md
├── LICENSE
├── README.md
├── configs  # Contains config files that control training parameters
├── data
│   └── benign_data
│       ├── benign_database  # Contains all the images
│       ├── coco_test.json  # Test annotations in COCO format
│       └── coco_train.json  # Train annotations in COCO format
├── detectron2_1
│   ├── __init__.py
│   └── datasets.py  # Registers train and test datasets
├── requirements.txt
└── train_net.py  # Main entry point for model training

Config

Each training run is completely defined by customizable parameters in its configuration file, with a few templates already specified in the configs folder.

For example, all the existing config files train the models with pretrained COCO weights:

• cascade_mask_rcnn.yaml: Cascade Mask R-CNN model with ResNet50 backbone.
• faster_rcnn.yaml: Faster R-CNN model with ResNet50 backbone.
• retinanet.yaml: RetinaNet model with ResNet50 backbone.

Other types of models and their respective configs and pretrained weights can be found in the official Detectron2 Model Zoo.

While you can refer to the config reference for a full list of available parameters and what they mean, I've annotated some of them in the existing configs, and some notable ones to customize are:

• SOLVER.IMS_PER_BATCH: Batch size
• SOLVER.BASE_LR: Base learning rate
• SOLVER.STEPS: The iteration number to decrease learning rate by GAMMA
• SOLVER.MAX_ITER: Total number of training iterations
• SOLVER.CHECKPOINT_PERIOD: Saves checkpoint every number of steps
• INPUT.MIN_SIZE_TRAIN: Image input sizes
• TEST.EVAL_PERIOD: The period (in terms of steps) to evaluate the model during training
• OUTPUT_DIR: Specify output directory to save checkpoints, logs, results etc.

Training

To train on a single gpu:

python train_net.py \
    --config-file configs/retinanet.yaml

To train on multiple gpus:

python train_net.py \
    --num-gpus 4 \
    --config-file configs/retinanet.yaml

To resume training from a checkpoint (finds last checkpoint from cfg.OUTPUT_DIR)

python train_net.py \
    --config-file config.yaml \  # Config file of halted run
    --resume

To see all options:

python train_net.py -h

Visualization

Training Examples

To visualize annotations of raw training images:

python visualize_data.py \
    --config-file config.yaml \  # Uses dataset specified in cfg.DATASETS.TRAIN
    --source annotation \
    --num-imgs 1000 \  # Num of images to save; don't specify if want to save all
    --output-dir /dir/to/save/images

To visualize annotations of training images after preprocessing/augs:

python visualize_data.py \
    --config-file config.yaml \  # Uses dataset specified in cfg.DATASETS.TRAIN
    --source dataloader \
    --num-imgs 1000 \  # Num of images to save; don't specify if want to save all
    --output-dir /dir/to/save/images

Predictions

To visualize JSON predictions:

python visualize_json_results.py \
    --input coco_instances_results.json \  # JSON file produced by model
    --output /dir/to/save/images \
    --dataset benign_test \  # Name of dataset
    --conf-threshold 0.5  # Default 0.5

Training Curves

Losses and metrics are automatically logged to TensorBoard.

tensorboard --logdir {path}

Evaluation

This command only runs evaluation on the test dataset:

python train_net.py \
    --eval-only \
    --config-file configs/retinanet.yaml \
    MODEL.WEIGHTS /path/to/checkpoint_file  # Path to trained checkpoint

Run DAG Algorithm

python run_DAG.py \
    --cfg-path config.yaml \
    --weights-path model_final.pth \
    --results-save-path coco_instances_results.json \
    --vis-save-dir saved